Workestral can do a lot more than a first glance suggests — from a chat assistant that takes action, to voice control, to standalone desktop and Mobile apps. A collection of practical tips.
The built-in chat assistant (✦ button, available on every page) knows the current workflow context and can carry out actions for you directly, instead of just explaining how to do it. A few examples you can simply type into the chat:
Use the history view (🕓) to switch between earlier conversations, "New conversation" (✏️) to start completely fresh, and "Pin view" (💾) to keep the chat window attached permanently — including its size.
Clicking 🎤 activates the microphone permanently — unlike voice assistants with a wake word ("Hey ..."), Workestral detects purely from volume when you start and stop speaking, and sends exactly that segment for transcription automatically. Click again to stop listening. Speech- to-text conversion runs locally (Whisper) — no audio data is sent to an external cloud service.
The speaker toggle is deliberately separate from the microphone: you can type text and have the reply read aloud, or speak via microphone without the reply being read back — whichever fits the moment. While a reply is being read aloud, the microphone is briefly muted automatically so the output isn't accidentally recorded again.
On the web, each team chooses its own active model under /speech-settings from
the Whisper models and Piper voices made available by the administrator, and can adjust
parameters like language, accuracy or speaking speed. In the Windows desktop app, this
choice is deliberately purely local to that machine (⚙ button next to the
chat) — independent of whatever the team has configured.
Instead of manually re-running a piece of research every day: enable an interval plus an optional fixed start time, so e.g. "daily at 07:00" really runs at the same time every day without drifting over the weeks. Workestral only emails you when something has actually changed — and every monitor also gets a fixed feed URL that can be subscribed to without email at all.
Write rules in plain language ("Automatically reply to support requests with an acknowledgement, but just flag invoices for review") — Workestral decides per email individually, not every message gets an automatic reply. For sending, it's worth setting up recurring subjects/texts once as a template instead of retyping them every time.
Every workflow step can use a different AI agent — e.g. research via a fast provider,
code review via a model that's particularly strong at code, image generation via a third.
An administrator sets this up via tags/priority on /agents.
Start for free: OpenRouter continuously offers free models. An administrator can click "Sync now" to automatically build a multi-tier fallback chain from the currently best free models — if the strongest tier fails (e.g. because a popular free model is currently overloaded), Workestral automatically tries the next-best one before a task actually fails.
If you ever see an error like "Could not retrieve a response": that usually means every single configured provider was unreachable at that moment (e.g. a shared rate limit) — just try again shortly after.
The Windows desktop app mirrors the same workflow as the web interface (task → refine → plan → result, including chat with microphone and voice output) and connects to the same Workestral server — runs and results are identical no matter which way you work. It's especially handy when:
Alongside the web and Windows desktop app, Workestral is also available as a native Mobile-App — with the same runs, the same chat, and the same note and planner management as on your computer. Handy when you're on the go:
The app is currently rolling out — contact us for early access.
On the plan page, before starting the workflow, "Add manual step" lets you insert a step of your own — with a short name, description and an assignment: either to a specific user, or to a functional role (e.g. "Finance"). With a role, any team member who has it can complete the step later, not just one person. The step's position can be adjusted with the arrow buttons — only for manual steps; automatically planned steps keep the order the router decided.
When the workflow reaches the manual step, it pauses and the assignee marks it done from the task page, optionally with a short note. The step then waits for approval by a different team member — anyone except the assignee themselves can approve. Before deciding, "Show output" opens an overlay with every result produced so far (finished reports/documents first, raw text output collapsed) — so the approval is based on the actual state of the work. If rejected, the step goes back to the assignee together with the reason.
Both the waiting assignee and every potential approver are notified by email (where an address is on file) and additionally see the open action prioritized at the top of the dashboard for as long as it's pending — a rejection or a restart triggers a fresh notification too.
If a paused workflow shouldn't continue after all, the creator or the currently responsible assignee/approver can cancel it with an explanatory note. The creator is notified by email and sees the note on the dashboard until they revise the workflow and restart it — which re-notifies every affected assignee.